Text Detection in Natural Scenes using Fully Convolutional DenseNets
Publish Year: 1397
نوع سند: مقاله کنفرانسی
زبان: English
View: 362
نسخه کامل این Paper ارائه نشده است و در دسترس نمی باشد
- Certificate
- من نویسنده این مقاله هستم
استخراج به نرم افزارهای پژوهشی:
شناسه ملی سند علمی:
SPIS04_030
تاریخ نمایه سازی: 16 اردیبهشت 1398
Abstract:
Text detection in natural scenes is challenging problem many researchers have worked on in the past decade. In this paper, we present novel approach to text detection using Fully Convolutional DenseNets. DenseNets have been very successful in object detection and recognition over the past couple of years, yet they have never been used to detect text before. To do text detection using FC-DenseNet, we train it to do semantic segmentation on the images. That is, we segment each image into sections that are text, background and word-fence. Using word-fence as segment allows for the model to learn in way that can separate close words. We use synthetic data to train our network from scratch, then use real augmented data to do fine-tuning. Our final model achieved the Recall and Precision of 0.83 and 0.71 on ICDAR 2013 dataset,respectively.
Authors